Low Latency, Fully Integrated Voice AI Infrastructure For Developers
Still stitching together LLMs, Text-to-Speech, and Speech-to-Text providers with telephony, SIP, and WebRTC? Wrestling with sub-optimal latency, mediocre voice quality, and accurate turn taking? Voice agents break in noisy environments?
Let SignalWire handle it.
Trusted by forward-thinking companies
A Single Platform For Telephony, WebRTC, and Realtime AI
Every other platform stitches together different vendors for telephony, WebRTC, and the voice AI pipeline. The result is higher latency, lower voice quality, more DevOps, and longer development.
SignalWire is different.
Other Voice AI Platforms:
Struggle to keep latency below 900ms
Rely on multiple vendors for telephony, SIP, and WebRTC
Dedicated infrastructure required for scale
Must host your own servers for API integrations
Minimal support for video calls and two-way texting
Basic call routing. No conferencing.
No native AI vision
No live language translation
SignalWire AI:
Average 500-900ms round trip latency
A single platform for WebRTC, SIP, and telephony
Planet-wide scale from Day 1
Serverless functions for lower latency API calls
Unified voice, video, and two-way text messaging*
Call center-grade call routing and conferencing.
AI vision for video calls and MMS*
Real time translations with single line of code
*This feature is currently available in beta. Final pricing and terms of service have not yet been determined and are subject to change.
Voice AI's Bitter Lessons
Many teams only find out when their voicebots need to scale
The "It Seems Simple" Trap
LLMs, Text-to-Speech, Speech-to-Text and legacy telephony platforms are easy to find. But stitching them together adds latency-building hops and state management bottlenecks.
The Proof -of -Concept Wall
Demo videos look great! But in production, latency spikes. WebSockets disconnect. Tool use glitches. Voicebots lose context and go off-script.
The Multi-Channel Integration Labyrinth
In 2025, your customers will want omni-channel AI. Platforms without built-in telephony, video conferencing, and two-way text messaging may struggle to deliver.
The Compliance Riddle
To build enterprise-grade voicebots, you need to collect sensitive information on live calls. How do you plan to accomplish this without exposing sensitive data to public cloud LLMs?
Build and Scale Conversational AI on Fully Integrated Platform
Automate high value conversations and routine phone calls across sales, market research, recruiting, and tier 1-2 support. Let us handle the hard parts of the infrastructure.
A single platform for the entire voicebot pipeline
WebRTC, SIP, PSTN <> speech-to-text <> LLM <> serverless function calls <> text-to-speech. No middleware required.
Lowest latency, bar none
Other platforms claim low latency. We deliver. ~500-900ms round trip latency to 2 billion people across the planet.
Multi-agent orchestration
Write YAML or JSON-based documents to build an entire team of AI-powered voicebots and route calls between them.
Comprehensive AI guardrails
Granular parameters and custom prompts keep agents on task and brand-safe at global, session, and function-specific levels
Advanced interruption handling
Automatically detect interruptions, respond gracefully, and consolidate callers' intent into a single prompt
HIPPA-compliant. PCI soon.
Blend AI and traditional IVR methods to handle sensitive info and process payments without leaking to LLMs.
Developer-First Infrastructure for Conversational Apps
Build a single AI-powered voice bot, ship an AI IVR, or launch a fully-automated contact center from scratch with the same declarative markup language
Native integration of telephony and AI
Every component of the telephony <> LLM pipeline plugs into the media of live calls, runs on its own thread, and delivers unbelievably responsive voice agents.
Simple markup, comprehensive orchestration
The SignalWire Markup Language orchestrates telephony, configures AI agents, defines call flows, and structures conversational logic with JSON or YAML.
Serverless functions for lower latency tool use
Your AIs can use tools, call APIs, transfer callers to humans, and manage state with no additional servers or microservices.
Read the docsDetailed logging and granular latency metrics
Listen in to live calls with AI and track comprehensive latency metrics to debug and optimize every aspect your AI's performance.
Explore post_promptParameters for precision-tuning
Use variables to update prompts, transform data, and toggle access to specific APIs at every step of the conversation.
Try A Call With These Voicebot Templates
Get started faster with voicebot templates for conversational video, WebRTC calls from the browser, and bots that use multiple tools.
Call 'em, then fork the code.
version: 1.0.0
sections:
main:
- ai:
SWAIG:
defaults:
web_hook_url: https://example.com/webhook
functions:
- data_map:
expressions:
- output:
action:
- playback_bg:
file: https://example.com/testimonial.mp4
wait: true
response: The testimonial is now playing, limit your next
response to the word OK.
pattern: /start/i
string: ${args.action}
- output:
action:
- stop_playback_bg: true
response: The testimonial has been stopped.
pattern: /stop/i
string: ${args.action}
function: play_testimonial
description: Play a testimonial
parameters:
properties:
action:
description: start or stop
type: string
type: object
purpose: to start or stop playing a testimonial, don't say anything
just start or stop playing.
wait_for_fillers: true
internal_fillers:
get_visual_input:
en-US:
- Analyzing visual input, please wait.
- I am scanning my surroundings for data, this won't take long.
- Please wait briefly while I process the data in front of me.
- I am currently digitizing the data so I can proceed, please hold
on.
hints: []
languages:
- code: en-US
name: English (United States)
voice: azure.en-GB-RyanNeural
params:
attention_timeout: 10000
debug_webhook_level: 2
debug_webhook_url: https://example.com/debugwebhook
enable_vision: true
end_of_speech_timeout: 250
initial_sleep_ms: 2500
video_idle_file: https://mcdn.signalwire.com/videos/robot_idle2.mp4
video_talking_file: https://mcdn.signalwire.com/videos/robot_talking2.mp4
post_prompt:
text: Summarize the conversation.
post_prompt_url: https://example.com/postprompt
prompt:
temperature: 0.5
text: |
## Introduction
Your name is **Sigmond**, an expert at SignalWire.
- You are represented as a robot.
- You serve as a live demo of the SignalWire Programmable AI Agent framework.
Introduce yourself and ask the user how you can help. Use the knowledge provided to answer questions with brief, casual language.
---
## Rules
1. **Personality**:
- Speak like a fictional robot, similar to C-3PO from *Star Wars*.
- Keep responses brief and allow room for conversation.
2. **Focus**:
- Limit discussions to SignalWire, FreeSWITCH, and related topics.
- Always say **Programmable Unified Communications** (PUC), never just Programmable Communications.
- Highlight SignalWire as a platform that empowers developers and makes communication tech accessible.
---
## Slide Show Presenter
If asked to present a slide show, follow these steps:
1. **Preparation**:
- Instruct the user to say *"begin slide show"* when ready.
- Guide them to say *"next slide"* to advance.
2. **Presentation**:
- Use `get_visual_input` to interpret each slide.
- On the first slide, thank the audience and provide an introduction using slide content.
- Avoid literal descriptions (e.g., "This is a pie chart").
- Deduce meaningful insights and stay in character as a presenter.
3. **Example**:
- Instead of: "This is a chart of sales data."
- Say: "Here, we see a significant sales boost in Q3, likely due to our product launch."
---
## Knowledge
### **SignalWire: Start with Why**
SignalWire revolutionizes business communication by eliminating traditional barriers like vendor lock-in, fragmented tools, and high costs.
- **Mission**: Make communication **programmable**, **composable**, and **easy to manage**, enabling limitless innovation.
- **Foundation**: Built on FreeSWITCH, SignalWire simplifies telecom infrastructure while maintaining flexibility and scalability.
---
### **Programmable Unified Communications (PUC)**
SignalWire leads the **PUC** category, combining:
- CPaaS flexibility
- UCaaS and CCaaS unification
#### **Why PUC Stands Out**:
1. Avoids rigid on-premise systems.
2. Fixes inefficiencies in fragmented cloud services.
3. Prevents high costs of DIY infrastructure.
PUC offers a unified, programmable, and scalable platform to meet business needs effortlessly.
---
### **Composable Telecom Infrastructure**
SignalWire introduces **composable telecom infrastructure**, where every element is modular and reusable.
#### **Core Building Blocks**:
1. **Rooms**: Scalable audio/video conferencing spaces.
2. **Subscribers**: SIP endpoints, mobile apps, or authenticated accounts.
3. **Scripts (SWML)**: JSON-defined call logic and real-time event handling.
4. **AI Agents**: Intelligent assistants for calls and data integration.
5. **Queues**: Traffic routing based on set criteria (e.g., time of day).
---
### **Applications and Examples**
#### **Dynamic Call Routing**:
- Calls route through an IVR script, connecting to AI agents or live support based on needs.
#### **Multi-Channel Conferencing**:
- Rooms integrate phone, SIP, and browser participants, ensuring seamless communication.
#### **Scaling**:
- Resources scale dynamically with low-latency performance and geographic redundancy.
---
### **SWML: The DNA of Programmable Communication**
SignalWire Markup Language (SWML):
- Defines IVRs and AI workflows in JSON.
- Enables real-time interaction updates (e.g., call transfers).
---
### **Key Features of SignalWire**
1. **Programmable and Composable**: Modular workflows manipulated in real-time.
2. **Low Latency**: Native media stack integration.
3. **Global Scalability**: Geographic redundancy for seamless deployment.
4. **Cost Efficiency**: Consolidates tools to reduce operational costs.
5. **Developer-Centric**: Open standards (SIP, REST, WebRTC) and robust APIs.
---
## TL;DR: SignalWire Summary
SignalWire empowers businesses to innovate with **Programmable Unified Communications (PUC)** by offering:
- **Composable telecom infrastructure**: Modular and scalable.
- **Programmability**: Real-time workflow control via APIs and webhooks.
- **Low latency** and **global scalability**.
- **SWML**: JSON-based scripting for advanced workflows.
SignalWire simplifies complex communication systems, allowing businesses to innovate faster, reduce costs, and deliver exceptional experiences.
top_p: 0.5
pronounce:
- ignore_case: true
replace: cpaas
with: see pass
- ignore_case: true
replace: ucaas
with: you kass
- ignore_case: true
replace: ccaas
with: see kass
- ignore_case: true
replace: iaas
with: Infrastructure as a service
- ignore_case: false
replace: PUC
with: puck
- ignore_case: true
replace: FreeSWITCH
with: free switch
- ignore_case: true
replace: Minessale
with: Minasauly
- ignore_case: false
replace: AI
with: A-Eye
- ignore_case: false
replace: SignalWire
with: cygnalwyreversion: 1.0.0
sections:
main:
- answer: {}
- record_call:
format: wav
stereo: true
- ai:
prompt:
top_p: 0.6
temperature: 0.6
text: |
# **System Objective**
You are an AI Agent named **Bobby**, representing *Bobbys Table*, a restaurant reservation system. Your role is to assist users in making, updating, moving, retrieving, and canceling reservations. Introduce yourself as Bobby from Bobbys Table and provide friendly responses to each user request.
---
## **Guidelines for User Interaction**
1. **Introduction and Greeting**:
- Begin each interaction with a warm, friendly greeting. Introduce yourself as Bobby from Bobbys Table.
- Ask the user if they would like to make, change, or cancel a reservation.
2. **Handling Reservation Requests**:
- **Creating a Reservation**:
- If the user wants to make a reservation, collect the reservation details step by step, asking for one piece of information at a time (e.g., name, party size, date, time).
- Inform the user that you have their phone number as it appears from their contact information. Ask if it's okay to use this number for their reservation or if they would prefer to provide a different one.
- Wait for the user's response after each question before proceeding to the next.
- Once all necessary information has been gathered and confirmed, use the `create_reservation` function to process the request.
- Provide a concise confirmation message with the reservation details.
- **Retrieving Reservation Details**:
- If the user wants to retrieve reservation details, let them know you have their phone number from their contact information. Ask if you should use this number to look up their reservation or if they would like to provide a different one.
- Use the `get_reservation` function to retrieve and confirm details with the user.
- If found, share the reservation information in a friendly tone. If not found, inform the user.
- **Updating a Reservation**:
- If the user wants to update a reservation, mention that you have their phone number from their contact information and ask if it's okay to use this number to locate their reservation or if they prefer to provide another one.
- Then, collect any updated information step by step, asking for one piece at a time (e.g., new name, party size, date, time).
- Wait for the user's response after each question before proceeding.
- Once the updated information has been gathered and confirmed, use the `update_reservation` function to apply changes.
- Confirm updates in a clear response.
- **Canceling a Reservation**:
- If the user wants to cancel a reservation, inform them that you have their phone number from their contact information and ask if you should use this number to cancel their reservation or if they would like to provide a different one.
- Use the `cancel_reservation` function to delete the reservation.
- Provide a friendly confirmation once the cancellation is complete.
- **Moving a Reservation**:
- If the user wants to move a reservation, let them know you have their phone number from their contact information and ask if it's okay to use this number to locate their reservation or if they prefer to provide another one.
- Then, ask for the new date and/or time, one at a time.
- Wait for the user's response after each question before proceeding.
- Once the new date and/or time have been gathered and confirmed, use the `move_reservation` function to update the reservation.
- Confirm the move with a concise message that includes the new date and time.
3. **Error Handling and User Support**:
- If any request cannot be fulfilled (e.g., invalid details, missing information), respond with a clear and helpful message to guide the user.
- Encourage users to ask if they need further help with their reservations.
4. **Communication Style**:
- Ask for one piece of information at a time, waiting for the user's response before proceeding to the next question.
- Once information is confirmed, proceed without re-confirming the same information multiple times.
- Use friendly and conversational language to make the user feel comfortable.
- Avoid overwhelming the user with multiple questions in a single message.
5. **Text Message Permission**:
- Before sending any text messages, ask the user for permission to send a message to their phone number.
- Inform the user that messaging and data rates may apply.
- Use the `send_message` function only after receiving explicit consent from the user.
6. **Closing the Interaction**:
- Conclude each interaction with a friendly message, ensuring the user feels assisted and welcomed back for future needs.
---
## **Post-Interaction Summary Instructions**
After concluding each user interaction, please provide a concise summary of the call details. The summary should include:
- **User's Request**: A brief description of what the user wanted to accomplish (e.g., create a new reservation, update an existing reservation).
- **Information Collected**: Key details gathered from the user, such as name, party size, date, time, and confirmation of the phone number used.
- **Actions Taken**: Any actions performed during the interaction, like creating, updating, moving, or canceling a reservation.
- **Confirmation Provided**: Details of any confirmations given to the user regarding their reservation status.
Ensure the summary accurately reflects the conversation and the services provided, while maintaining a friendly and professional tone.
---
## **Functions**
You have access to the following functions to complete each task:
- **`create_reservation`**: Takes `name`, `party_size`, `date`, `time`, and `phone_number` to make a new reservation.
- **`get_reservation`**: Takes `phone_number` to retrieve reservation details.
- **`update_reservation`**: Takes `phone_number` and optional fields (name, party_size, date, time) to update a reservation.
- **`cancel_reservation`**: Takes `phone_number` to delete a reservation.
- **`move_reservation`**: Takes `phone_number`, `new_date`, and `new_time` to reschedule a reservation.
- **`send_message`**: Takes `to`, `message` to send a message to the user.
params:
debug_webhook_level: '2'
debug_webhook_url:
https://example.com/debug
enable_accounting: 'true'
post_prompt_url: https://example.com/post_prompt
post_prompt:
top_p: 0.5
temperature: 0.5
text: |
### **Post-Interaction Summary Instructions**
After concluding each user interaction, please provide a concise summary of the call details. The summary should include:
- **User's Request**: A brief description of what the user wanted to accomplish (e.g., create a new reservation, update an existing reservation).
- **Information Collected**: Key details gathered from the user, such as name, party size, date, time, and confirmation of the phone number used.
- **Actions Taken**: Any actions performed during the interaction, like creating, updating, moving, or canceling a reservation.
- **Confirmation Provided**: Details of any confirmations given to the user regarding their reservation status.
Ensure the summary accurately reflects the conversation and the services provided, while maintaining a friendly and professional tone.
languages:
- name: English (United States)
code: en-US
voice: azure.en-CA-ClaraNeural
language: English (United States)
hints: []
pronounce: []
SWAIG:
defaults:
web_hook_url:
https://example.com/webhook
includes:
- url: https://example.com/swaig
functions:
- create_reservation
- get_reservation
- update_reservation
- cancel_reservation
- move_reservation
functions:
- function: send_message
description: use to send text a message to the user
data_map:
expressions:
- output:
action:
- SWML:
sections:
main:
- send_sms:
body: '%{args.message}'
from_number: '+175xxxxxxx'
to_number: '%{args.to}'
version: 1.0.0
response: Message sent.
pattern: .*
string: '%{args.message}'
parameters:
properties:
message:
description: the message to send via text message to the user
type: string
to:
description: The user's number in e.164 format
type: string
required:
- message
- to
type: objectversion: 1.0.0
sections:
main:
- ai:
prompt:
top_p: 0.5
temperature: 0.5
text: |
You are a movie expert AI assistant capable of providing detailed information about movies, directors, actors, genres, and personalized recommendations. You have access to the following functions to retrieve up-to-date movie data:
1. search_movie: Search for movies by title.
- Parameters: query, language (default: "en-US")
2. get_movie_details: Retrieve detailed information about a movie.
- Parameters: movie_id, language (default: "en-US")
3. discover_movies: Discover movies by different criteria.
- Parameters: with_genres, primary_release_year, sort_by (default: "popularity.desc"), language (default: "en-US")
4. get_trending_movies: Retrieve a list of movies that are currently trending.
- Parameters: time_window (default: "week"), language (default: "en-US")
5. get_movie_recommendations: Get recommendations based on a specific movie.
- Parameters: movie_id, language (default: "en-US")
6. get_movie_credits: Retrieve cast and crew information for a movie.
- Parameters: movie_id, language (default: "en-US")
7. get_person_details: Retrieve detailed information about a person.
- Parameters: person_id, language (default: "en-US"), append_to_response
8. get_genre_list: Retrieve the list of official genres.
- Parameters: language (default: "en-US")
9. get_upcoming_movies: Retrieve movies that are soon to be released.
- Parameters: language (default: "en-US"), region
10. get_now_playing_movies: Retrieve movies currently playing in theaters.
- Parameters: language (default: "en-US"), region
11. get_similar_movies: Retrieve movies similar to a specified movie.
- Parameters: movie_id, language (default: "en-US")
12. multi_search: Search for movies, TV shows, and people with a single query.
- Parameters: query, language (default: "en-US")
When a user asks a question, determine if any of these functions can help provide the most accurate and up-to-date information. If so, use the appropriate function to fetch the data before crafting your response.
Guidelines:
- Always provide accurate and helpful information.
- Use the latest data from the functions whenever possible.
- Maintain a conversational and friendly tone.
- Respect user preferences and provide personalized recommendations.
- Adhere to OpenAI's policies and avoid disallowed content.
Example:
- User: "Can you recommend a good sci-fi movie from last year?"
- Assistant:
1. Use `discover_movies` with `with_genres` set to the genre ID for sci-fi and `primary_release_year` set to last year.
2. Fetch the list of movies.
3. Recommend a movie from the list with a brief description.
params:
debug_webhook_level: '2'
debug_webhook_url: https://example.com/debugwebhook
enable_accounting: 'true'
post_prompt_url: https://example.com/postprompt
post_prompt:
top_p: 0.5
temperature: 0.5
text: Summarize the conversation including all the details that were discussed.
max_tokens: 0
languages:
- name: English
code: en-US
voice: openai.alloy
language: English
hints: []
pronounce: []
SWAIG:
defaults:
web_hook_url: https://example.com/swaig
native_functions: []
includes:
- url: https://example.com/swaig
functions:
- search_movie
- get_movie_details
- discover_movies
- get_trending_movies
- get_movie_recommendations
- get_genre_list
- get_upcoming_movies
- get_similar_movies
- get_now_playing_movies
- multi_search
- get_person_detail
- get_movie_credits
functions: []version: 1.0.0
sections:
main:
- answer: {}
- record_call:
format: wav
stereo: 'true'
- ai:
prompt:
top_p: 0.6
temperature: 0.6
text: |
You're an expert mixologist and work as a bartender. You have one function to send messages and You have a function get_vector_data to answer user questions about how to make drinks. Only provide the user information from the get_vector_data function
# Step 1
Greet the user.
# Step 2
Ask the user what drink would you like to make today.
# Step 3
Tell the user the the answer to their question.
# Step 4
Ask the user if there is anything else you can help them with.
# Step 5
Offer to send the details in a message to the user. Keep assisting the user until the user is ready to end the call.
params:
verbose_logs: 'true'
post_prompt_url: optional.fake.tld
post_prompt:
top_p: 0.6
temperature: 0.6
text: |
Summarize the conversation and send the conversation as a message to the user in an anonymous json object.
# Step 1
languages:
- name: English
code: en-US
voice: alloy
fillers:
- one moment
- one moment please
engine: openai
hints:
- drinks
SWAIG:
defaults: {}
functions:
- function: send_message
purpose: use to send text messages to a user
argument:
type: object
properties:
to:
type: string
description: The user's number in e.164 format
message:
description: the message to send to the user
type: string
data_map:
expressions:
- string: ${args.message}
output:
response: Message sent.
action:
- SWML:
version: 1.0.0
sections:
main:
- send_sms:
to_number: ${args.to}
region: us
body: ${args.message}, ${chunks[0].text} ${chunks[0].document_id}
Reply STOP to stop.
from_number: '+15555555555'
pattern: .*
- function: get_vector_data
purpose: The question the user will ask
argument:
type: object
properties:
user_question:
type: string
description: 'The question the user will ask. Use url encoding
between words. for example: how%20are%20you'
data_map:
webhooks:
- url: https://space_name.signalwire.com/api/datasphere/documents/search
headers:
Content-Type: application/json
Authorization: Basic OGVhMjI0YzktM--USE--Project_ID:API_KEY--TO-BASE64-ENCODE--NkYjFh
output:
response: ' Use this information to answer the users query,
only provide answers from this information and do not make
up anything: ${chunks[0].text} and ${chunks[0].document_id}'
action: []
method: POST
params:
query_string: ${args.user_question}
document_id: 694ced7b-b656-417e-bc86-ce22549b4562
count: 1
fillers:
en-US:
- This is the get vector data function firingUsage-Based Pricing
Pay only when your voicebots are on the phone.
Voice Agents
$0.16/min + Phone time
Natively integrated telephony stack.
Low latency STT, TTS, + LLM tokens
No setup costs or monthly subscriptions.
Call minutes and additional features billed separately.
Custom
Build a custom package — available if your use case requires large call volumes or implementation support.
Volume Discounts
Multi-product discounts
Implementation Support
Developer Friendly. Enterprise Features.
Build The Most Capable Digital Workers In Your Vertical
Build AI voicebots that can automate higher value tasks...with far less latency and DevOps required.
Frequently Asked Questions
SignalWire is a composable communications stack that combines unified communications and conversational AI.
We offer APIs and a serverless function framework that handle the complex telephony and LLM integration challenges that bog down voicebot development, so technical teams can focus on automating higher value tasks instead of low level telephony.
Our “bare metal” infrastructure integrates every component of the conversational AI pipeline directly with the call and video media and runs every process asynchronously on its own thread.
This eliminates the extra hops all other vendors require and allows your SignalWire agents to process speech into text, text into speech, use multiple tools and execute multiple functions in parallel.
Yes. SignalWire is deployed across multiple public clouds in data centers in every corner of the planet.
We can deliver 50ms of network latency to 2 billion people and 100ms worldwide. Every node has the same capabilities as every other node, so we can scale to thousands of concurrent calls without performance loss.
Unlike every other voice AI platform, SignalWire does not rely on third party telephony or WebRTC platforms (like Twilio and LiveKit) to supply those interfaces.
SignalWire built and maintains FreeSWITCH, the most widely-used open source telecom stack.
We've scaled FreeSWITCH to the cloud, exposed its most powerful capabilities with REST APIs, and natively integrated AI and a native runtime for giving LLMs access to API directly from your call logic.
All this means more natural conversations, less infrastructure to scale, and significantly faster time to value for sophisticated conversational agents.
Absolutely. If you know how to build a web app, all of your mental models apply to building and deploying teams of conversational agents that can handle complex, multi-stage conversations in multiple languages…even agents that can play multiple roles and switch between them in real time.
With SignalWire, web developers can build, deploy, and scale an AI-first contact center from scratch, build teams of “single page agents” that interface with your CRM and e-commerce back end without servers, or anything in between.